Moonshot AI Releases Kimi K3: Open-Weight 2.8T MoE Model

LinkMoonshot AIAnnouncement

Moonshot AI has released Kimi K3, an open-weight native multimodal agentic model featuring 2.8 trillion total parameters and a 1-million-token context window. The architecture uses a sparse MoE setup that activates 104 billion parameters per token.

Key points

  1. Total size of 2.8T parameters using Mixture-of-Experts (MoE), routing 16 out of 896 experts (104B parameters) per token alongside 2 shared experts.
  2. Built on 69 Kimi Delta Attention (KDA) and 24 Gated MLA layers with Attention Residuals (AttnRes) and a 401M-parameter MoonViT-V2 vision encoder.
  3. Context length of 1,048,576 tokens with native understanding of text, images, and video.
  4. Trained with quantization awareness featuring native MXFP4 weights and MXFP8 activations.
  5. Released under the Kimi K3 License with deployment instructions for Transformers, vLLM, SGLang, and Docker.

Why it matters

Kimi K3 marks the availability of an open 3T-class model targeting long-horizon coding and agentic workflows, competing on standard benchmarks against frontier systems like Claude Fable 5 and GPT-5.6 Sol.

Analysis

Moonshot AI focuses heavily on sparsity: combining LatentMoE with KDA yields an estimated 2.5x scaling efficiency boost over Kimi K2. By keeping active parameters down to 104B and relying on native MXFP4 quantization, the 2.8T architecture aims to manage serving and compute requirements.

Mentioned links

Found in

Summary generated by AI (Gemini 3.8 Flash, 8 October 2026). It may contain errors – the original source is authoritative.

Related entries

Similar in content, found by the AI search.

  • Repository:PipeNetwork/kimi-k3-mlx

    kimi-k3-mlx: Apple Silicon Port of the 2.78T Multimodal MoE Kimi-K3

    kimi-k3-mlx is an Apple MLX port of Moonshot AI's native multimodal MoE model Kimi-K3 (2.78T total parameters). Because the model exceeds the memory capacity of any single Apple Silicon machine, the repository includes a custom streaming converter, vision components, and expert-pruning tools.

    337starsPython

    AI & LLMs· Tool

  • Repository:studio-dots-ai/dots3-note-prev

    dots3-note preview: Open-Weight 280B MoE Model with 512K Context

    Dots Studio has released dots3-note preview, the first open-weight model in the dots3 family. The Mixture-of-Experts architecture features 280 billion total parameters with 16 billion activated parameters, multimodal inputs, and context support up to 512K tokens.

    123stars

    AI & LLMs· Announcement

  • Link:Xiaomi MiMo Team

    Xiaomi Releases MiMo-V2.5-Pro as Open Source

    Xiaomi has unveiled MiMo-V2.5-Pro, an open-source Mixture-of-Experts model featuring 1.02 trillion total parameters with 42 billion active. The model supports a 1M-token context window and is specifically designed for complex agentic workflows and long-horizon software engineering.

    AI & LLMs· Announcement

  • Link:ornith-ai

    Ornith-1.5-397B: Open-Source MoE Model Focused on Coding and Reasoning

    ornith-ai has released Ornith-1.5-397B, an open-source multimodal mixture-of-experts model under the MIT licence. With 397 billion parameters and an end-to-end self-improvement loop, it delivers coding and reasoning benchmark performance comparable to top proprietary models.

    AI & LLMs· Resource

  • Link:Z.ai

    Z.ai Releases GLM-5.3-Flash as an Open-Weights Multimodal Model

    Z.ai has launched GLM-5.3-Flash under the MIT licence on Hugging Face, marking the first natively multimodal model in the GLM-5 family. Featuring 320B total parameters and 18B active parameters, it targets high efficiency across coding and agentic tasks.

    AI & LLMs· Tool

Back to the knowledge base

Let's talk about your project

Get in touch

Locations

  • Mattersburg
    Johann Nepomuk Bergerstraße 7/2/14
    7210 Mattersburg, Austria
  • Vienna
    Ungargasse 64-66/3/404
    1030 Wien, Austria

Parts of this content were created with the assistance of AI.